← Back to blog
Grok 4.7 vs 4.6: What Changed and Should You Upgrade?

Grok 4.7 vs 4.6: What Changed and Should You Upgrade?

Julian Brooks

By Julian Brooks

MyClaw Editorial

Compare in MyClaw

Grok 4.7 vs Grok 4.6

Run the same hard job in both models to compare quality, recovery, speed, and cost before choosing a stack. Neither is an automatic default.

AI Takeaway

  • Which model should you choose? Start with Grok 4.7 for new, demanding coding or agent workflows. Its published gains justify testing it first; keep 4.6 where your existing results remain better.
  • Is Grok 4.7 more expensive? Both start at $2 per million input tokens and $6 per million output tokens. Token consumption, retries, tools, and service tiers determine the actual bill.
  • Did the context window grow? Both xAI API model pages list 500,000 tokens. Platform limits and billing thresholds still matter.
  • Should you keep Grok 4.6? Keep it where it meets your quality and budget targets until 4.7 demonstrates a useful improvement.

Grok 4.7 vs 4.6 at a Glance

Grok 4.7 launched on September 21, 2026, following Grok 4.6 on August 12. xAI describes a larger base model trained longer on difficult tasks. The core API specifications remain the same:

SpecificationGrok 4.7Grok 4.6
API model IDgrok-4.7grok-4.6
Release dateSeptember 21, 2026August 12, 2026
Base input / output per 1M tokens$2 / $6$2 / $6
Base cached input per 1M tokens$0.50$0.50
API context window500,000 tokens500,000 tokens
Input → outputText and images → textText and images → text
Reasoning effortLow, medium, high, xhighLow, medium, high, xhigh
Default effortHighHigh

Prices reflect xAI’s global API rates for prompts below 200K tokens, checked September 22, 2026. Both support function calling and structured outputs. Our Grok 4.6 review covers the previous release in detail.

Your application still determines which tools the model can use. A chat that suggests a code fix and an agent that edits files and runs tests give you different ways to judge the same model.

Introducing Grok 4.7 | SpaceXAI

What Changes for Coding and Long-Running Work?

Coding and Terminal Tasks

xAI’s September 21 launch comparison shows gains across these three benchmarks:

xAI-reported benchmarkGrok 4.7Grok 4.6
CursorBench 4.046.3%40.4%
DeepSWE v1.171.0%65.2%
Terminal-Bench 4.038.0%20.3%

Effort settings differ: the table compares 4.7 at xhigh with 4.6 at high, except the 4.7 DeepSWE result, which is marked high. These are vendor-published results, not our independent tests.

Terminal-Bench rises by 17.7 percentage points, while DeepSWE gains 5.8 points at the same named effort level. These results support trying 4.7 on harder engineering tasks; they do not establish equal compute use or predict your success rate.

For coding-agent workflows, judge the patch itself. Did the agent locate the cause, preserve unrelated behavior, and run a test that could catch the original bug? A confident final message is weak evidence if the repository still fails.

Watch what happens after a failed command. Does the agent diagnose the error and change approach, or repeat the same attempt? That recovery behavior can matter more than how quickly the first edit appears.

Grok 4.6 for Agentic Coding and Knowledge Work

Research, Documents, and Self-Checking

xAI also reports better self-verification and document work. A useful test is whether 4.7 catches mistakes that would otherwise reach your final deliverable.

For example, give both models the same source pack and request a short briefing. Check whether every conclusion follows from a source, conflicting figures are reconciled, and missing information stays visibly unresolved. For a spreadsheet, inspect formulas and totals; for slides, compare claims against the supplied material.

In a recurring research-agent workflow, track unsupported claims and missing evidence across several reports. A 500K context window leaves room for substantial source material, but does not guarantee accurate retrieval or reasoning throughout it.

Grok 4.7 vs 4.6 Pricing: What Will the Task Cost?

Shared Base Rates and Platform Differences

The base rates above apply to short-context requests. The current xAI pricing table places both models in the higher tier when a prompt reaches 200,000 tokens: $4 input, $1 cached input, and $12 output per million tokens. Those rates apply to all tokens in that request.

Regional routing also matters: xAI lists a 10% premium for its US regional endpoint. Cursor and other providers may apply different allowances, thresholds, or charges; check the platform you actually use.

The 4.7 launch also describes an optional fast variant at twice the output speed and twice the price. Keep that separate from a standard-model comparison.

Compare Cost per Completed Task

Consider a hypothetical short-context call at global base rates:

  • 100,000 uncached input tokens plus 10,000 billed output tokens cost $0.26.
  • The same input plus 20,000 billed output tokens costs $0.32.

These are illustrative calculations, not measured model usage. They exclude tools, hosting, caching, and surcharges. Identical token prices can produce different bills when the amount of work changes.

Use this practical metric across your test batch:

Cost per accepted task = total model and tool spend across all attempts ÷ accepted tasks.

Define an accepted task as one that meets your requirements. Include failed attempts and repairs in the spending total, and track human review time separately. If every attempt fails, neither model has delivered a usable result, regardless of how little the batch cost.

Our guide to managing OpenClaw token costs covers context and caching considerations for recurring work. Keep those settings consistent so you can distinguish model behavior from savings caused by a warmer cache or shorter prompt.

Measure completion time independently. Output speed tells you how quickly text arrives; extra reasoning, tool calls, and retries determine how long the whole job takes.

Compare in MyClaw
Grok 4.7 vs Grok 4.6

Run the same hard job in both models to compare quality, recovery, speed, and cost before choosing a stack. Neither is an automatic default.

Should You Upgrade to Grok 4.7 or Keep 4.6?

Test 4.7 on Work That Stalls

For a new demanding workflow, 4.7 is our recommended starting candidate. For an existing one, begin with tasks where 4.6 needs repeated repairs or leaves work unfinished. Those failures give the upgrade a concrete problem to solve.

Prioritize the outcome that matters to you. A coding team may value fewer broken patches; an analyst may value fewer unsupported claims. Decide how much extra cost or waiting time you would accept for that improvement before inspecting the answers.

Adopt 4.7 when the improvement matters: more accepted results, less cleanup, faster completion, or better cost per finished task.

Keep 4.6 Where It Already Delivers

An established workflow does not need an immediate replacement. If 4.6 already meets its targets, retain it while testing 4.7 alongside it.

For short summaries, routine rewrites, or simple questions, the long-task benchmarks provide less direct guidance. If both outputs meet your requirements, response time and actual charges may settle the choice.

For agent testing, keep files, tools, starting states, and prompts fixed. Start at matched effort, then assess xhigh separately. The same effort label does not guarantee the same compute or token consumption. Repeat runs and set a spending limit so one expensive success does not hide a pattern of failures.

Keep a way to return to 4.6 while your provider offers it. You can adopt 4.7 for difficult jobs and retain 4.6 for tasks where it still performs well.

Try Grok 4.7 vs 4.6 in MyClaw

Use MyClaw Chat to try Grok 4.7 and Grok 4.6 on your own work. Bring a code snippet, source material, or a draft you need improved, then compare how close each answer gets to a usable result.

1. Bring One Real Task

Choose one clear assignment and define what a good answer must include. For example, paste a function with a known bug and ask for a fix plus a regression test, or provide source material and request a brief with traceable conclusions.

Include constraints such as length, format, and facts the answer must preserve. Reuse those instructions in both chats.

Get Started

2. Run It in Both Model Chats

Use the buttons below to open the Grok 4.7 and Grok 4.6 chats. Start a fresh conversation in each and submit the same task, context, and output requirements. Keep any available settings consistent.

Let each model finish before adding follow-up instructions. If either answer needs a correction, save that exchange so the extra effort counts in your comparison.

3. Compare the Results and Choose

Compare correctness, completeness, and how much editing each answer needs. Run suggested code or tests before accepting a fix. Check response time and usage information where available, and consult MyClaw Chat plans for current allowances. Chat subscriptions are billed separately from API credits and agent hosting.

Start with one task today. If the results are close, try a harder example from the same workflow. Let the quality of the finished work decide which model you use next.

Compare in MyClaw

Grok 4.7 vs Grok 4.6

Run the same hard job in both models to compare quality, recovery, speed, and cost before choosing a stack. Neither is an automatic default.